1 /* Copyright (c) 2013-2017. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "../colls_private.h"
8 //#include <star-reduction.c>
11 // NP pow of 2 for now
12 int Coll_allreduce_rab1::allreduce(void *sbuff, void *rbuff,
13 int count, MPI_Datatype dtype,
14 MPI_Op op, MPI_Comm comm)
18 int tag = COLL_TAG_ALLREDUCE, send_size, newcnt, share;
19 unsigned int pof2 = 1, mask;
20 int send_idx, recv_idx, dst, send_cnt, recv_cnt;
24 int rank = comm->rank();
25 unsigned int nprocs = comm->size();
27 if((nprocs&(nprocs-1)))
28 THROWF(arg_error,0, "allreduce rab1 algorithm can't be used with non power of two number of processes ! ");
30 extent = dtype->get_extent();
33 while (pof2 <= nprocs)
37 send_idx = recv_idx = 0;
40 if ((count % nprocs)) {
41 send_size = (count + nprocs) / nprocs;
42 newcnt = send_size * nprocs;
44 recv = (void *) smpi_get_tmp_recvbuffer(extent * newcnt);
45 tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * newcnt);
46 memcpy(recv, sbuff, extent * count);
50 share = newcnt / pof2;
53 send_cnt = recv_cnt = newcnt / (pof2 / mask);
56 send_idx = recv_idx + (mask * share);
58 recv_idx = send_idx + (mask * share);
60 Request::sendrecv((char *) recv + send_idx * extent, send_cnt, dtype, dst, tag,
61 tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
63 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) recv + recv_idx * extent, &recv_cnt,
66 // update send_idx for next iteration
71 memcpy(tmp_buf, (char *) recv + recv_idx * extent, recv_cnt * extent);
72 Colls::allgather(tmp_buf, recv_cnt, dtype, recv, recv_cnt, dtype, comm);
74 memcpy(rbuff, recv, count * extent);
75 smpi_free_tmp_buffer(recv);
76 smpi_free_tmp_buffer(tmp_buf);
81 tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * count);
82 memcpy(rbuff, sbuff, count * extent);
87 send_cnt = recv_cnt = count / (pof2 / mask);
90 send_idx = recv_idx + (mask * share);
92 recv_idx = send_idx + (mask * share);
94 Request::sendrecv((char *) rbuff + send_idx * extent, send_cnt, dtype, dst,
95 tag, tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
97 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) rbuff + recv_idx * extent, &recv_cnt,
100 // update send_idx for next iteration
105 memcpy(tmp_buf, (char *) rbuff + recv_idx * extent, recv_cnt * extent);
106 Colls::allgather(tmp_buf, recv_cnt, dtype, rbuff, recv_cnt, dtype, comm);
107 smpi_free_tmp_buffer(tmp_buf);